PARSIMONIOUS MODEUNG OF YIELD CURVESFOR U.S. TREASURY BILLS
by
Charles R. Nelson and Andrew F. Siegel
TECHNICAL REPORT No. 60
February 1985
Department of Statistics
University of Ifashington
Seattle. Washington 98195
PARSIMONIOUS MODELING OF YIELD CURVES
FOR U.S. TREASURY BILLS
By
Charles R. Nelson and Andrew F. Siegel
Universitv of WashingtonSeattle. Washington 98195
Februarv 1985
Not to be Quoted without permission of the authors
The authors wish to thank the Center for the Study of hanKIng andFinancial Markets at the UnIyersity of Washington for supporting thisresearch. Nelson also received support from the National 5cienceFounOatlOn under a grant to the National Bureau of Economic Research whichis acknowledged with thanks. Research assistance was provided b FrederickJoutz and Ann Kremer. We are grateful to Vance Roley for obtaining thedata set used in this st dy.
Abstract
A new model is proposed for representing the term to maturitv
structure of interest rates at a point in time. The model Droouces humoed.
monotonic, and S-shaped yield curves using four parameters. ConditIonal on
a time decaY parameter~ estimates of the other three are obtained bv least
squares. Yield curves for thirty-seven sets of U.S. Treasurv bill vields
with maturities up to one vear are presented. The median standard
deviation of fit is Just over seven basis points and the corresponding
median R-squared is .96. Study of residuals suggests the existence of
specific maturitv effects not previously identified. Using the models to
predict the price of a long term bond provides a diaonostic check and
suggests directions for further research.
1. Introduction
The idea that there is a svstematic relationshio between vield and
term to maturity on debt instruments is a oersuaSIve one and accounts tor
one Ot the largest lIteratures in monetarY economics. that of the term to
maturity structure of interest rates. On a purelv descriptive level. the
scatter of points recording observed vield and term to maturitv for
securities within a particular class at a given time strongly suggests the
existence of an underlying smooth function relating yield t~.w maturitv.
Such a function is called a vield curve.
The fitting of yield curves to vield/maturity data g02S back at least
to the pioneering efforts of David Durand (1942) whose method of fitting
was to position a French curve on the scatter of points in such a wav that
the resulting curve appeared subjectivelv reasonable. Yield may be
transformed to present value and J. Huston McCulloch (1971. 1975) has
proposed apprOXImating the present value functIon by a pieCEwise polynomial
spline fitted to price data. Garv Shea (1982~ 1984) has shown that the
resulting yield function tends to bend sharplv towards the end of the
maturity range observed in the sample. This would seem to be a most
unlikely property of a true yielo curve relatIonshIp and also suggests that
these models would not be useful for prediction outside the sample maturity
range. Other researchers have fItted a varietv of parametric models to
yield curves, including Cohen~ Kramer. and Waugh {1966J. Fisher \1966!;
Echols and Elliott (1976J. Dobson (1978); and Chambers. Carleton and
Waldman (1984 Some of these are based on pol nomial regreSSIon. a d a1
i elude at least a linear term which would force extrapolated verv 10ng-
term rates to be unboundedlv larce \either positive or negative! despite
theIr abIlities to fIt closelv wIthin the range of the data.
Fong (1982) have suggested exponential solines as an alternative to
polynomial splines. In a comparison of the two splIne methodologies~ Shea
(1983) finds that exponential splines are subject to the same shortcomings
that POlVnOmlal splines are.
That there is a need for readilv implemented techniques tor fItting
vield curves seej~to us apoarent from the popularitv of vielo curves as a
tool of analvsis in financial markets. Market letters from major brokerage
houses, government publications, and even the ~€~ lQCE Iim~§ cater to
readers interest in seeing a representation of the underlying relationshIp
between vield and maturitv by publishing graphs of yield curves. To our
knowledge, these are fitted bv free hand methods. We feel that it ought to
be possible to develop a computer-based method for calculating and plotting
yield curves in real time WhICh is both more satisfactory from a conceptual
vieWpOInt than are polynomial splines and less dependent on the judgement
of an Individual observer than is free hand sketching.
The objective of this paper is to present the prototvpe of a
parSImonious modeling procedure which we believe meets these objectives.
We have tested the procedure on U.S. Treasury bill YIelds taken from quote
sheets at four week intervals over a three vear period. The algebraic form
of the model is motivated bv the solution function for a second order
dIfferential equatIon and generates humped~ monotonIC and S-shaped cur es
USIng four parameters. We find t at the model fits t e oi 1 vield data
wit a media standard de lat o ust over seven baSIS points and
prodUCES a media R-square or about .96. A 1 three basic ield curve
shaDes are encountered in the sample. Studv of the residuals reveals
specifIC maturItv effects not prevIouslv identified. Extrapolation of
yields outside the maturity range of bills allows us to predict the prIce
of a long term bond. Comparison of the actual bond price with that
predicted bv the basic model sugoests refinements to the fitting procedure
and directions for further research.
2. Motivation for the Model
A U.S. Treasurv bill is a promise to pay the amount of its face value
on a stated maturItv date. Since there are no Interim coupon payments on a
bill the market price is necessarily less than its face value. The yield
on the bill is defined to be that rate of return which produces the face
value from an investment equal to the market price in the time remaining
until maturity. Arbitrage assures that all bills with a giyen maturity
date sell at the same price. and therefore have the same vield. at anv
instant in time. Bills of different maturities may of course sell at
prices which imply different yields to maturity at the same point in time.
The yields on any two bills of dIfferent maturities imply a forward vield
or rate for the time interval between the maturities of the two bllls. If
the maturities are l say, ml
days and ffi2
days (m2
j mIl then an investor can
secure the forward rate of return for an 1m2
- mIl
day perIod to begIn ffi1
days hence b selling hills of ffi, davs to maturitv and replacing them with~ !' ,
bills of ffi_ days to maturity.£
The incentive to do this will vary directl
with the difference between the forward rate available in the market and
the inyestor s assessment of the rates of return which are likel to be
available in the arket on bills of m2-
1) davs to maturit m1
da s from
the present. ThIS suggests that expectations of future bill yields
influence the term to maturity structure of yIelds observed In the market.
It also suggests that forward rates will not e hIblt increasing
fluctuations as one considers longer maturitIes because it seems
Implausible that expected future interest rates would vary increasinglY as
one looks further into the future.
Considerations of this sort lead us to posit that a satisfactorY model
for the yield curye must imply forward rates that are smooth as a function
of horizon and that oscillations in the function. It anv. must damo down.
These will also be properties of the vield curve because yield to maturitv
can oe exoressed as a smoothing of the intervening forward rates.
SpecificallY. consider the forward rate implied by bills of m days to
maturity and im +~! days where ~ IS arbitrarily small. This is an
instantaneous forward rate which we will denote bv rim;. The definitIon of
the forward rate implies that
R',ml = lim f: r(}:ldx
where Rim) is the YIeld to maturitv on a bill maturing in m days. Thus.
yield to maturity is just an average of the forward rates.
the forward rate rim) is given by
rim) = Hind + mR (ml
Equivalently.
where R (m) is the slope of the yield curve at maturit m. This second
eq ation pOInts out that an wrInkles in the yield curye. giVIng rise to
la ge values of t e slope!
consider larger maturities.
a e a magnifIEd effect on or ward rates as we
If our hypot esis that the forward rate
functIon becomes smoother with increasing m is correct. then the relation
between R m) and m must be eyen smoother.
Do actual vlelds on bills! olotted agaInst maturity. display the
smoothness we expect to find) To form a prelIminary impression, consider
the plot of U.S. Treasurv bIll yields displayed in Figure 1. These are
continuouslY comnounded yields at an annual rate computed from closing
asked dIscount yields on the New York Federal Reserve quote sheet for
January 1981. The yields rise as a function of maturitv until about
100 days maturity and then decline generally until about 300 days maturity
where theY appear to leyel off. To be an acceptable candidate to fIt this
data a function needs to have the caDability of rising to a maximum and
then fallIng monotonically towards an asYmptotic yalue.
Polynomials are clearly not acceptable by thIS criterion. While thev
readily form hump shapes theY do not settle down to an asymptotic value but
instead head off towards plus or minus infinity. Of course, bv choosing a
polynomial of su~ficiently high degree we can get a very close fit to the
data in the sense of generating a curve which comes close to the data
points. A polvnomial of degree equal to the number of data points less one
can be constructed that coincides exactly with each data point. The
function itself will fluctuate wildly between the data points and one would
have to be willing to believe that yields on bills are coinCIdentally close
to one another onlv at the specific maturities which the Treasury happens
to have issued. This essential difficulty with the behavior of polvnomials
can be mitigating b the met od of spJi es which uses low degree
po ynolllials to It different sectl s of t e maturit spectr Ill, Oil 9
them together at points called knots. It is eas to imagIne that the data
F e c se e
the rance zero to 130 days. another over 130 to 180 days! and a thlro over
18 to 350 days. ThIS would be in the spirit of McCulloch S work although
he fitted splInes to prices rather than yields. Our view of pOlynOmial
splines is that they are a patchwork approach to the problem which does not
overcome the fundamental ShortcomIng of polynoffiials~ that theIr slope tends
to increase "(absolutelvl towards longer maturities. From the relation of
yields to forward rates it is clear that the forward rate function will
dIverge eyen more rapidly. Extensive analysis of spline results bv Shea
\1982, 19841 shows that even when the fitted yield curve appears reasonable
within much of the maturitv range of the data! the implied forward rates
disolav erratic behavior at the high end of the range. We would like to
develop a class of models which incorporate intrinsically the smoothness
and asvmptotic damping we expect ~ QC~QC! of yield curves and the Implied
forward rates. Such models could meaningfullv address the question: what
is the vieid we may expect to see on a 360 day bill to be issued by the
Treasury today. given observed yield on bills presently tradIng which have
maturities only up to 330 days/ In contrast, polynomial splInes are poorly
equipped to predict out-of-sample.
A class of models which does possess the properties we seek is that
formed bv the solutions to ordinary differential or difference equations.
Since the latter will be more familiar to most readers, consider the second
order difference equation
r m) = a1
rim-I) + u_ r m-2) +L
a d the e alua ion of r m) for m=1.2!3, ..• gIven i itial alues r a d
r(-II. The dynamiC behavior of r m) will of course deDend on the alues of
,«1 and "2 through the characteristic equation l-U
1X-U2XL
= 0. while
[«~j1-«1-a~] WIll be the asymptotic level of rim) as m gets large if thev ~
eouation is stable. If the roots are real and lie outside the unIt circle
the solution has the form
+ B_ exol-m/T_)L' L
where i 1 and 1 2 are positive constants determined by "I and «2~ 81 and 82
are constants determined bv the initial conditions. The oarameters 1> and1
1~ are time constants which determIne the rate at WhICh the terms exp\-m/llL
decay to zero. Thus , at maturitv m=l we have expl-l) = 0.37~ at maturIty
m=2i we have expl-2) = 0.14, and so forth. As m gets laroe both
exponential terms become small so rim) approaches ao as its asYmptotic
level. Differing rates of decay implied bv 11
and T allow rim) to take on2
humped shapes as well as monotonic shapes.
Some theories of the term structure of interest rates imply forward
rate equations of this form. The classical expectations theory equates
forecasts of short rates l which might be represented bv a stochastic
difference equatIon. to forward rates. Richard (1978) studied a model in
which the term structure depends on two state variables: the real rate and
the rate of inflatIon. Under certaintv the forward rate functIon In
Richard s model has precisely the above form if the two state variables
each obey a first order differential equatIon. Under uncertaintv the
forward rate is a more complex function of exponentials. While we do not
fee obliged to tie ou model to any specifIC model of the term str cture!
these considerations add to the presuffiotion that thIS is a class of models
worth InvestigatIng.
Imolementation of thIS model presents some oractical difficultIes
val es of those paIrs of parameters are sWItched around we have the same
functIon, a potential source of computer confusion. A more readllv
Imolemented model with similar shape characterIstIcs has the torm
(2. 1 rim) = 8~ + 8 exoi-m/t} + B_[ mittilL
ThIS functIon arIses as the solution to the second order dIfference
equation in the case of equal roots. or alternatIvely may be derlved as an
approximation to the solution in the unequal roots case bv replacIng one of
the two exponentIal terms bv its Tavlor s series expansion IAppendix A).
The parameters of this model are more easilv estImated because the model IS
linear in B.• Bland B~. f.cr any pr ov i s i ona I val ue of T.t) L
Model (2.1) may also be viewed as a constant plus a laouerre Function.
which suggests a method for generalization to higher-order models.
Laguerre Functions consist of a pOlYnQmlal times an exponential decay term
and are a mathematical class of approximating functions; details may be
found (for example) in Courant and Hilbert 11953, pp , 92,-97J.
While higher order models could generate more cample shapes, it is
not hard to show that even the second order model given above has
considerable shape flexibilitv and IS therefore parSImonIOUS. Note that
r(OJ is \8.+B1
) and the limiting value of rem} as m gets large is simply BtI
Setting theSE arbItraril at zero and one respectlvelv for the purpose 0
st ding shape and noting again that IS onl a time sea e arameter and
ma be set at on for t e same urposE, we are left WIt a f ction of one
parameter onlY
rifl'lI:::: l - (i-am; e>:pi-mJ.
Allowing this single shape parameter to vary from -6 to 12 in equal
increments oroduces the range of shapes seen in Fioure 2. These include
humpsl S-shaped. and monotonic curves. Shapes produced bv vertically
inverting these curves are also oossible under this model. easily allowing
decreasing curves.
To obtain vield as a function of maturity for the second order model
one integrates r(*)
resulting function is
in i2.1) from zero to m and divides bv m. The
(2.2) fdmJ :::: B.. + i81
+ B~) .. Ll r ex p i-miT) JI imiTJ - BL~ ex p i-miT)t) .i..
which is also linear In coeffiCIents. given 7. The limIting value of Rim)
as m gets large is BO and as m gets small is (BO+8 1/which are necessarily
the same as for the forward rate function SInce Rim! is Just an averaging
of r 0:). The range of shapes available for Rim; depends again on a single
parameter SInce for '{ :::: 1.. B. ::::o i , and <B. + til)U
:::: (I we have
Rim) :::: 1 - (1-a) * [l-expi-mJJ!m - a * end-mi.
Allowing a to take on values from -6 to 12 in equal Increments generates
the shapes displayed In Figure ~ which include humps. S-shapes. and
monotonic curVES. On the basis of the range of shapes available to us In
the second order model our operating hypothesis is that we will be able to
capture the underlying relationship between yield and term to maturitv
WIt out resorting to more comole>: models invo vlng mare parameters. Wood
1983 presents ield curves fitted b traditio a1 me hods an ual from
1900 through 1982 and allot them fall within the range of generic shapes
which can be generated by our model.
Another way to see the shape tIe ibilitv of the second order model is
to rearrange its elements into long term and short term components as
tallows
The expression in braces may be interpreted as the long term component of
the yield function because it gives the only rearrangement of terms which
starts out at zero and also decays at a rate much slower than exponential,
namely 11m. The last term is the short term component since it starts at a
value ot unity and has the fastest possible iexDonentialJ decay to zero.
This decomposition is illustrated in Figure 4. It is easy to see now with
appropriate choice of weights tor these components we can generate curves
with humps in them and ones which are monotonIC ibut not necessarIly a
simple exponential function.)
3. Empirical Yield Curves forU.S. Treasury Bills
The objective of our empirical work is to assess the adequacy of the
secono order model for descrIbing the relatIonship between yield and term
to maturity for U.S. Treasurv bills. The data come from Federal Reserve
Bank of New York quote sheets sampled on every fourth Thursdav \excepting
holidays) from Januar 22. 1981 through October 27. 1983. thirt -seven in
all. The quote sheets give the bId and asked dIscount and bond eq iYalent
yield tor the bills in each maturitv date outstanding as of t e close at
tradIng on the date 0 the quo e sheet. Number of da s to rna urlt is
calculated from the delivery date, which 15 the followIng Mondav tor a
Thursdav transaction, until the maturIty date. Typically there are thirty
two maturitIes traded, which on these Thursdays work out to terms of from 3
days to 178 days In increments of seven days, 199 davs, and then increments
of 28 days to 339 davs. On three dates there was also a one year bill
traded. T~e bid and asked discounts have been calculated as if there were
a 360 daY year and are on a simple interest basis. The bond equivalent
yield is Intended to present the bill yield on a basis comparable to that
of a bond which pays a half-yearly coupon. The exact formula for doing
this is not, to our knowledge, available publically. Bill prices
themselves are not displayed but are readily calculated from the discount
yields. We have converted the asked discount to the corresponding price
(that paid by an investor) and then calculated the continuously compounded
rate of return from delivery date to maturity date annualized to a 365.25
day year. These yields are the data we fit to the vield curve model.
Observations on the first two maturities, 3 and 10 days, are omitted
because the yields are consistently higher, presumablY due to relatively
large transaction costs over a short term to maturity. This leaves thirty
yield/maturity pairs observed on each of thirty-four market dates and
thirty-one on three dates.
For purposes of fitting vIeld curves we have oarameterized the model
(2.21 in the form
i3.1) Riml = a + b * [l-exp -m/i)] I (m/i) + c * exp(-m/i .
For an proviSIonal value of i we may readily calculate sample values of
the two regressors. The best fitti 9 values of the coefficients a, b~ and
c are then computed using linear least squares. Repeating this procedure
o er a range at values tor i reveals the overall best-fitting values of 1,
a, b, and c. Recall that I is a time constant which determines the rate at
whIch the regressor variables decav to zero. Plots Ot the data sets reveals
that the vield/maturitv relationship becomes quite tlat in the range 200 to
300 davs las in Figure Ii, suggesting that best-fitting values of' would
be in the range 50 to 100. We consequently search over a grid from 10 to
200 In increments of 10, and also 250, 300, and 365.
Small values of , correspond to rapid decay In the regressors and
therefore will be able to fit curvature at low maturities well, while beIng
unable to fit excessive curvature over longer maturity
Correspondingly, large values of 1 produce slow decav in tne regressors
which can fit curvature over longer maturity ranges but will be unable to
follow extreme curvature at short maturities. This trade-oft is
illustrated in Figure 5 which shows the vlelds observed on Februarv 19.
1981. The yields rise qUIte sharply at low maturities, from 13.80 percent
at 17 days to 14.94 percent at 59 davs maturity. This portion of the data
is fitted much better by a model with 1 = 20 than one with 1 = 100 as shown
bv the two continuous curves plotted in Figure 5. On the other hand. the
The best overall fit for
smaller 1 value produces a poor fit over the maturity range above 200 days
relative to that provided by the larger 1 valUE.
this data set is given by T = 40 {not plottedi.
It is also quite clear from Figure 5 that no set of values of the
parameters wall d fit t e data perfec ly, nor is it our objecti e to find a
model which wou d do so. A more highly parameterized model WhICh could
follow all the wiggles in the data is less likely to predict well! in our
view, than a more parsimoneous model which assumes more smoothness in t e
underlving relatIonshIp than one observes in the data. There are a number
ot reasons why we would not e pect the data to conform to the ~[~~
underlying relationship between vield and maturity even If we knew what it
was. For example, there is not continuous trading in all bills. so
published quotes will reflect transactions which occurred at different
pOInts in tIme during the trading day. Bills of specific maturities may
sell at a discount or premium. We hope that by studYing departures of the
data trom the fitted model we can Identifv §y~!g~~!t£ as well as
idosyncratic features of the data which the model is failing to capture.
The basic results for the second order model fitted to each of the 37
data sets are presented in Table 1. The first column gives the data set
number, the second column the best fitting value of I, the third column the
standard deviation of residuals in basis points (hundreths of a percent),
and the fourth column the value of R-squared. Median values of these
statistics over the 37 samples are given at the end of the table. The
first point worth noting is that the model accounts tor a very large
fraction of the variation in bill yields~ median R-squared is .959. The
median standard deviation of residuals is 7.25 basis points, or .0725
percentage points, Dr a .000725 in yield. Standard deviations range from
about 2 baSIS points to about 20. Best fitting values of 1 have a median
of 50. Thev occurred at the lower boundar v of the search range \1=10) In
two cases and at t e upper boundar v 1=365) in three cases. The first data
set, which was seen in Fig re is dlspla ed in Figure 6 along wit the
itted yield curve. It is clear from t e oattern of deviations from the
curve that residuals are not random but rather seem to e hibIt some
SetNo.
TABLE 1
Second Order Hodel5td. De. R- Std. [Ie.
at best 1 50 0 1=5
First Term OnlyStd. Dev,at best
1 5(1
,,' 4~ 30.'4 60c- 40J
6 407 808 10 " 1 )
9 :LI,)
1,', :.0')
1, 30.1
12 30013 5014 3015 601o 10 ( 1 )
17 11018 2019 17020 21::21 2022 365 { 1 J
23 4024 3025 2026 100-'7 300" 1
28 5029 11t)
..)\.1 7o31 365 ( 1 )
32 2033 4034 12035 9036 365 ( 1 )
37 ,80..
Median 50
16. 09 92. 4 16.09 46. 7113. o« 88. '7' L.b! 36. 4211 .-:;"") 72. "' 12. 45 13. 46LL .'
6. 01 86. I b •, ") 0 0('.. L j ·
12. Co",:. 87. 8 14. r-j 30. 97I ... ~l ...
13. 47 ..., " 3 1:!1,52 13. ~ -,, ...i. "-·k
15. 61 49. ! 15. 90 17. i 110. n 81 7 22.42 23.0(:19. 85 88. 8 LO. 34 1°
c;-,! • ~b
1B. 77 95. 2 18. 1":: 18. H',...1.:,. '-' '-'
4. 80 98. 8 6. 11 6 . 95'u
1"') 28 93. 8 12. 43 12. 16....! • 76 99. 4 7. 76 7·67
11.08 98.0 1 1.32 11 ~ -.i....t.
10. c;-, 95. 7 1(I.-c;- 15. 20..!1 !..J
6. 28 97. 3 7. .:;.\) 7.55c: 11 98. '7 c- 71 5. 74"'. .' .J.
7 .51 86. 4 10. 12 11·1(1
4. 12 98. B 4.46 4. 055. 79 98. 8 9.26 9.98
20. 04 96. 7 "le:- 17 it:' re:-LI.,J. ....101 • .....tl-l
15. 08 98. , 15. 84 15.41.'10. 01 99. 1 11. 'r= 14. 78O..J
"') 91 99. 6 e:- 13 6. 17.... ~'.7 'jt:' 97. 4 7. 45 7. 34, .... ..!
r 18 93. 9 r=~,..) 5.09..!. ..J.
3. 71 97. 3 4. 03 3. 655. 38 95. 5 5. 38 5.286. 72 85. 6 6. 90 6. 591.95 98. (1 i. • iO 2.213. 74 'I ' 6 4. 02 ~" 68, 1 .4. 89 96. 1 5. 80 4.833. 16 9° 1 3.22 .) .19. !.
7.24 96. 1 7 82 7. 1115. 34 OL " 15.51 15. 07\,JW • .'
5. 5' 95. 9 6. 17 5. 43.'3. 01 99. 0 4. 25 '1 97.;. .
7.25 qr: 9 7. 82t o ,
NOTES: best fit realized at boundar of range of search.Standard deviations are i basis points.
dependence along the maturity aXIS. We therefore refrain from makIng
statements about the statistIcal sIgnifIcance of coeffIcient estimates
based on conventIonal standard errors. We wIll also be interested to see
If such patterns are svstematIc across samples.
A small value of I will be indicated in cases where the vields change
sharply at low maturities and then level oft quicklY as in the case of data
set No. 8 for August 6~ 1981 plotted in Figure 7 along with the fItteo
yield curve for I = 10. Slow curvature which decays slowlY will be fit
best by a large value of 1 as in the case of set No. 22 for September 2.
1982 plotted in FIgure 8 with the fitted vield curve for 1 = 365. What 15
not readIlY apparent in Figure 8 is that the plotted portion of the curve
represents only the rising portion of a verv long hump \see Figure 9) WhICh
ultimatelv decays to an asymptotic vield of -.025. Clearlv the best fit to
the sample does not guarantee sensible extrapolation. Althougn the best
fitting valUES of I varv considerablv. as these examples indicate. rather
lIttle preciSIon of fit is lost if we impose the median value of 50 for 1
for all data sets. The resulting standard errors appear In the fiith
column of Table 1 and have a median value of 7.82 basis points! or onlY .57
basis points higher than when each data set was allowed to choose its own
I. For a few data sets this constraint makes a noticeable difference. as
in the case of data set No.8! for example! a small 1 seems preferable.
However In the cases where I was 365 the constraint costs lIttle In terms
of preCIsion. The ave all results s ggest that little may be gaIned in
practIce bv fItting 1 to eac data set IndIvId ali
The lowest alue of R-souared recorded was 49. for set No. 7 while
t e hIghest was 9.0 for set No. 24.
5
The characterIstIcs of the two data
sets which lead to thIS results are evident in FIGUreS Ivana 11
respectIvelv. Data set No. 7 in Figure appears to be two data sets at
dIfferent levels which a smooth curve will have little abilIty to account
for. This apoarent dIscontInuity is rare in our sample and may reflect
lack of late tradIno in the long sector of the market that day! or perhaps
clerical er r. In contrast! data set No. 24 in Figure 11 presents a verv
smooth. S-shaped pattern WhICh IS very precisely tracked bv the model
leaving resIduals with a standard deviation of onlv about 3 basis points.
The abIlity of the second order model to generate hump shaDes was one
of its attractIve attributes conceptually, but the question remains whether
this fleXIbility is important empirically. An alternative more simple
model would be a simple exponential function for forward rates obtained by
setting 8~ equal to zero in equation (2.11. The corresponding vield
function then has onlv the fIrst term! in which maturity appears in the
denominator! but not the second term as can be seen by setting 8~ = 0 inL
equatIon (2.21. Only monotonic yIeld curves can be generated bv thIS
restricted model. The final column of Table shows the standard
deviatIons or residuals resultIng from Imposing this constraint (but now
allowing 1 to take its best fitting value). The median over the 37 data
sets IS 9.00 basis pOInts comoared with the 7.25 reported for the
unconstrained model. In some cases the standard deviation rises sharolv.
For example~ it is no surprise that a monotonIC curve does not fIt the
first data set well; the standard deviation rIses from 16.09 to 46.71 basis
points. In some cases the standa d deviation is reduced slightl because
t e constrai ed model fits about as well and uses one less a ameter. e
abi itv to fit umps seems to have been Quite Importan until the twenty-
t da se Ja ua 98 er
the shape of the vield curve seems to have become simpler and monotonic.
h s there appears to be a persistence of shape over time. Note that this
change In shape also seems to be associated wIth less disperSion In
residuals. Did the Federal Reserve start to stabilize Interest rates again
In late 1982'1 A casual InsDection of the behavior of the federal funds
rate over t~IS period certainly suggests that it dlO.
4. Analysis of Residuals: Maturity and Issue Effects
Plots of fitted yield curves against the oat a have suggested some
dependence of residuals along the maturity aXIs. We would like to try to
determine whether this IS due to a systematic Influence of maturity on
vleld which our model is unable to capture. If such an effect persists
through time then we should be able to detect it In the average of the
thirtv-seven residuals corresponding to a specific maturity. Figure 12 is
a vertical stack of reSidual plots for the thirty-seven data sets with the
averaged residual at the bottom. The individual residual plots are
separated bv intervals of 200 basis pOints and the scale for the averaged
reSIduals 15 magnified by ten. The last data set appears at the top Ot the
stack. Note that the first averaged reSidual, correspondIng to 17 days
maturitv, is positive, the second negative, followed by a rising pattern to
Just under nInety davs, a sharp droD, then a riSing pattern again to just
under 180 davs and another sharp drop. This is seen more clearly in Figure
13 where the magnified scale shows that these maturity effects are in the
ra DE -5 to +5 baSIS poi ts whic I s large relative to a rou standard
de Ia ion of ~, oaSIS pOints. We surmIse that the POSitIve YIeld effect• I-
at 17 days is due to i 9 er transaction costs per unit ti me for shorter
term bills. The titted curve is pulled upward bv this data point, leaving
tne ne t point below the curve. We also surmIse that the peak at 87 Days
maturIty and sharp drop tollowlng 15 due to the fact that 90 davs is the
maturity of a SUbstantial aortion at the bills issued bv the rreasur ana
Will therefore bulk laroe in the inventorv on dealers shelves. Simllarlv,
the rreasurv issues 180 dav bIlls and 360 dav bills and Indeed we observe
the averaged resIdual riSIng to a peak at each of these maturitIes. To our
knowledge, these supplY effects have not been prevIously documentea nor
would they be apparent It our models did not impose quite a bit of
smoothness on the vield curve. A ourchaser of bills mav or may not find
these maturity premiums sufficlentlv attractive to influence maturity
choice, but at least they are now VISIble.
Issue effects are distinguished from maturity effects in that they
pertain to the bills Which mature on a particular date rather tnan to bills
WIth a particular term to maturity. The issue of bills maturing on
December 31. 1981 were 339-day bIlls on our first quote sheet (Januarv
1981). became 311-day bIlls on our second quote sheet (February 19, 1981)
28 days later, and so on through the months until they appear as 29-dav
bIlls on the November 27. 1981 quote sheet. This oives us twelve reSIduals
for thIS particular issue of bills. Other issues will appear initially
with only 17B-days to maturity which gives us six residuals until the Issue
matures and disappears. The plots of residuals are lined up in Figure 14
so that each issue may be followed through time. Averages are plotted at
the bottom with the scale enlarged by a factor of three~ these a erages are
s 0 on a larger scale in FIgure 15. There is some eVIdence In these
pots that issue effects exist since large residuals for a partIe lar Issue
show some tendencY to persIst from one quote sheet to the oe t. For
e ample. the Issues due January 1982 and Januarv 14, 1982 exhibiteo
large negatIve resIduals In the nInth data set \September 3, 1981 and did
agaIn a month later In the tenth data set (October 1. 19811. nowever no
abnormal deVIation was eVIdent thereafter. SImIlarlY, the Issue due
September 1982 was assocIated witn a large negative resIdual In the
twentv-first data set (Hugust 5, 1982i and agaIn in the twenty-second.
after which it came to maturity. Eyidence for issue effects is less
compelling than that for maturity effects but would seem to warrant further
investigation.
5. Prediction Out-of-Sample:Pricing a Long Term Bond
One of our criterIa for a satIsfactorv vleld curve model is that It be
able to predict yields beyond the maturity rangs of the samole used to fit
it. An unreasonably exactIng test would be to ask it to predict the vield
or prIce of a long term government bond. but this IS what we have tried to
do. The particular bond chosen IS the 12-3/4 percent coupon U.S. Treasurv
bond maturing in 2010 lcallable In 2005) since this was the longest bond
appearing on all our quote sheets. A bond Lan of course be viewed as a
bundle of bills with maturIties spaced at six month intervals until the
maturity date of the bond. Each component bIll pays an amount equal to
the semi-annual coupon except the last which also pays the face value of
the bond. Values read off a yield curve can be used to discount each
component bi 1 in the stream and the resulti 9 tota value can be compared
WIt the quoted rIce Ot t e and, adjusting first for accrued interest
from t e last co pon date WhICh the b yer must pa to the se Ie •
The predicted bond price will Ot course deoend primarily on the
portion at the yield curve which lies beyond the range of the sample bill
data because at most onlv the first two semi-annual coupon paYments can be
due within the one year maturitv limit Ot U.S. Treasurv bills. For our
yield curve model with values of 1 around 50 the fitted curve flattens out
considerably for maturities beyond a Year. The first exponential term In
the model goes from unltv at zero maturity to .1369 at 365 days maturity,
and the second term goes from unitv to .0007 in the same interval. The
priCing of the bond is therefore determined largely by the asymptotic level
of the curve oiven by the intercept in the model, EquivalentlY, the
value of the intercept must be close to the vield to maturity on the bond
if the model is to price the bond accurately. Figure 16 is a plot of the
actual price of the bona chronologically for the thirty-seven dates In our
sample (light line! and the corresponding predicted prices (dark line!
produced by the model when we allow 1 to take its best fitting value. Two
predictions are drastically awry, the twelfth i$138.063 against an actual
price of $100.34) and the twenty-second ($404.58 against an actual prIce Ot
i>103.59i. These were both models which had large values of 1 (see Table
2i. In both cases the bill yield data was fitted as the rising portion of
a long hump with eventual decay to a much lower level which was .079 for
the twelfth model and. as the reader may recall, ~~Q§~i~@ .v25 for the
twenty-second. Tne resultino discount rates are therefore too low and the
predicted dond orice correspondingly too high. Constraining to i3 value
of 5 in both cases costs little in standard deviatIon of fit see Table
but improves t e p edicti 5 of bond p ices dramaticall and
ill 2.52 r e sp ec t i el The improvement is evident in Figure were t e
predicted bono prices have been generated from models fitted under the
constraint that' is 50 Ithe median value of 1 across the samples).
The relation between actual and predicted bond prIce also is depicted
as a scatter plot in Figure 18. It is obVIOUS that the correlation between
actual and predicted price is high. nume~lcallv it is .963. but also that
the predictions overshoot the actuals. The magnitude of overshooting IS
much larger than could be accQunted for by favorable tax treatment for the
bond when It is selling at a discount from face value. This suggests that
our fitted curves may flatten out too rapidly. When yields generally were
high and the vield curve downward sloping the models overestimated longer
term discount rates and therefore underestimated the price of the bond, and
the reverse was true when yields were relatively low and the yield curve
was upward sloping. Correcting the price predictions for these systematic
biases by simple linear regression! we obtain a standard deviation for the
adjusted bond price prediction of only 12.63. Evidently, the value of i is
best chosen bv fitting across data sets rather than by selecting the value
for each individual data set.
What correspondence is there between the ability of a model to fit the
bill yield data well and its accuracy in extrapolating beyond the sample to
predict the yield on a bond? The short answer is: none necessarily. A
function may have the flexibility to fit data over a specific Interval but
have very poor properties when extrapolated outside that interval. A cubic
polynomial has the same number of parameters as does our model and indeed
fits the bill vield data slig tly better. T e median standard deVIation of
residuals is anI .1 baSIS points over the thirty-seven data sets.
However we know that a cubic polynomIal in maturlt Will head off to eIther
L
plus infinity or minus Infinity as maturity increases, the sign depending
on the sIgn of the CUbIC term. It is clear then that if we use a cubic
polynomIal yield curve to prIce out a bond it wIll assign eIther very great
present value or very lIttle present value to dIstantly future payments.
For our data set the result IS predicted bond prices which bunch in the
intervals il1 to 140 and $384 to $408. The correlation between actual and
predicted bond price is -0.020, so the polynomIal model has no predictIve
value although it fits the sample data very well.
b. Summary and Conclusions
The solution function of a secone order differential equatIon orovides
the baSIS for a parsimoneous mOdel capable of representIng the range of
shapes prevlouslv associated with the YIeld/term to maturity relationship
or vieJd curve. It has a number of properties which are appealIng a
Q[!Q[!: smoothness: ability to assume monotonic, humped and S-shapes: ana
asymptotic dampIng. The model IS able to account for aaout 90 percent of
the variation in U.S. Treasury bills across maturities during the 1981-83
sample periOd WIth a standard deviatIon of resIdual errors of 7.25 oasis
points. Analysis of residuals reveals maturIty effects which seem to be
related to the specifIC maturities issued by the Treasury. Extrapolation
of the fitted curves to price a long term Treasury bond suggests that the
oasic tIme constant In the model exhibIts conSIstency over time and is best
chosen on the baSIS of average experience across data sets rat er than
Indl lduall for each ield curve. Given an appropriate value for t 2 tIme
c stant the th eE remaining parameters are fitted b si Ie least sq ares,
making the proceoure operational in real time. H polynomial fIts t 2 bIll
12 a aa a as we ed c 5 00 -0 -sa
APPENDIX A: THE PROPOSED MODEL AS AN APPROXIMATIONIN THE UNEQUAL ROOTS CASE
The pr co os ed model 12.1! 1 S
'. A. 1 ! r vm .
Our purpose is to show that this solution in the case of two equal roots of
the characteristic equation is in fact an approximation to the unequal root
solution when those roots are not very different.
Suppose the two roots give rise to decay rates 1. and 1_ and hence to1 L
the model
riml = i + t eyp'-m'l I +(I 1 ,. , I i: L e:,:pt-11l1'L).1.. J-
If we write I/'r = 1/'1 ~ I1/T~~ L
1/11 and exoand part of the second1
exponentIal term to first order in this difference~ we find
rim! "'I + ¥, e}:pl-mI'(1)
+¥_[e);ol-m/'!,I)ll-ml'T1-T_I!'l.'T_.1t) 1 i.' 1 1..1:":
~ ''i. + (Y +1?!01.;.
WhICh we recognize as beIng In the form of our proDosed model IA.l) with a
suitable reparametrization.
REFERENCES
Chambers, Donald R.~ Carleton, Willard T.; and Waldman, Donald W.Approach to the Estimation of the Term Structure of Interest~Q~[Q~l Qi Etn~Q£l~l eDg @~~D~11~~lY~ eD~lY~12 I.IX (Septemberpp.
'A NewRates .:
1984) •
Cohen. Kalman J.; Kramer, Robert L.; and Waugh, W. Howard. 'RegressionYield Curves for U.S. Government Securities,' ~§~~g~~~Q1 §£l§Q£§ 1.111(December, 1966), B-168 to B-175.
Courant, R.,(1953) •
and Hilbert D.WIley: New 'ork.
Volume 1
Dobson, Steven W. 'Estimating Term Structure Equations with IndividualBond Data,' ~QY[D!! Qi Eln!~£! AXXIII (March, 1978), 75-92.
Durand, David. ~§?l£ It!lQ? Q± ~Q[QQ[§1! ~QnQ2~ 1~~Q:!!1~, National Bureauof Economic Research, Technical Paper 3. 1942.
Echols. Michael E., and Elliott. Jan Walter. 'A Quantitative Yield CurveModel for Estimating the Term Structure of Interest Rates,' ~QY[n§l Q!tio§o£t~! IOQ ~Y~D1!t!ilY! ~Q§iY§l? XI (March. 19701, 87-114.
Fisher. Douglas. 'Expectations, the Term Structure of Interest Rates. anaRecent British Experience,' ~£QnQ~l~! XXXIII (August~ 19661, 319-329.
McCulloch, J. Huston. 'Measuring the Term Structure Ot Interest Rates,'~QYrne! Q! ~!d2t!]~?? J,LIV (Januarv, H71i, 19-31.
'The Tax-Adjusted Yield Curve, ~Q~[!]~l Q± Ei!]§!]£§ xXX {June.19751,811-829.
Richard, Scott F.An Arbitrage Model of the Term Structure of InterestRates, ~Q~[!]21 Q! Einln£!!l ~£QnQID1~§ VI (19781, pp. 33-57.
Shea, Gary S. 'The Japanese Term Structure of Interest Rates. unpublishedPh.D dissertation, University of Washington, 1982.
'Interest RateWorking Paper,(November 19831.
Term Structure Estimation with Exponential Splines,Board of Governors of the Federal Reserve SYstem
Pitfalls in Smoothing Interest Rate Term structure Data: EquilibriumModels and Sp ine ApproprIations, ~QY[!]~l Q! EiUe!]£l~l
~Q~i1~ilY~ ~lr?l? XIX \5eptember 1984 , pp.
Wood, Joh H.I terest~~Qt gi
Do leld Curves Normall Slope eRates, 1862 1982,' ~£Q!) s §Q~~UY~§~
1£~9Q (Ju y!August 1983 • 17-23.
erm Str CtU e ofF §[~! ~~§![Y§
-
o•.....
·CD
o•
xxxxxxxx
xxxx
x xxx
x
xx
xx
x
Yieldo 0• •...... ......U1 0)
x
x
x
x
x
•oo
•.....W
0+---+----f----I--01---+---+---I----I--1--I
......
~oo
cTW'<0
o
......DOo 0
'<(j)
("t'
o I\.)
3 0o 0cTC.,
E)co("t'
lD(j)
.....<,I\.)I\.)
"CD.....•
r»;........I
0'-'
* Model curve.....,........ 0I
rnx-0r»;I
3--I'-/
L..J ....,.
"" 33 mI c-r
a-0
3 Ul
*rn 0x c-r-0 C,........ )
I....,.
3 c-r<:» '<
fllX-0
••
Model curves,...,f-A 0rt=::::j===::!::::!:::::=;::::!:==:::±:=--r--t--t
rnX-0
" -II3 ......'-' 3i-J ill<,
'(1'"30
3 Ul
rn 0X (1'"
-0 Cr»; "1I ......
3 (1'"
<:» '<
£)
c0(1" Yieldm(J)
0 0 0 0 0• • • • •
r\.) ........ ........ ........ ....... .......<,r\.) w ~ U1 m........
(0 0<,00.......•
(1"
0 ........C DO
0 0II '<
(J)
ro(1"0 a f\.)
03 0
:J 00.. 0
(1"
........ c0 -,0
.....,.• (1"
'<
O. 16
O. 14
>- O. 15
400100 200 300Days to maturity
0.13 -I I 1 I f 1 1 1- I
o
.,....,
lJr-t
OJ
1..9
O. 18tJIt.
5II'")
, J
LLO. 17
Quotes 1/22/81, tau = 50.
£)Yield
c 0 0 0 0 00 • • • • •("T ....... ....... ....... ....... .......ID ~ U1 0) .......:J .0)(J)
00)<, x(J) x
x<,co....... .........
0 00 0
("T '<0 (J)c
("'tII 0
....... 3 x0 0• ("T x
CJ......("T x
'<x
x
£) Yieldc 0 0 0 0 0 0 0 0a • • • • • • • •c-r 0 0 0 0 0 f-oI' f-oI' f-oI'ID(J) U1 en '-J co to 0 f-oI' f\.)
0to<,f\.)<,cof\.)
f-oI'..DO
c-r 0 00 "'<c (J)
II c-ra
w·3Q)
U1 0• c-r
c.,......c-r
'<
~oo
0Yield
c 0 0 0 0 00 • • • • •("t". ......,. ....... ....... ....... .......ID W ~ Ul 0) '-J(J)
0'-J x<,to<,co....... ......,... DO
0 0(1'" '<0 Q)c
("t"II 0 f\)
co 3 00 0 0• ("t"
c.,.......rt' W
'< 00
0 Yieldca 0 0 0 0 0('1" • • • • •(0 0 0 0 0 ~
(j) en ""-J CD (Q 00
~
0<,I'\)CD<,CD ~
I'\) 00• 0 0'<
('1" (j)0c ("'t
a1/
3W 00 ('1"• c
)......('1"
-c
:::um(J)......0..CQI-'(J)
Q:Ja.Q
<m"1Q
illro"1m(J)......0Co........•
o rrrr-rrrn1J
~
DOQ 0
'<(J)
("to3 0Q(1"C"1......(1"
'<
AVQragQ rQsidual yieldt I I
0 0 0 0 0 0 0• • • • • • •0 0 0 0 0 0 D·0 0 0 0 0 0 0....... ........ a a 0 ........ .......UI 0 Ul a UI a U1
a
.......00c 0
'<(f)
(T
o roa
3 0o(Tc-,......(TW'<0
o
'I iI
II
I
-r ~
I I I I I I I I I I I I III I II I I I I I' I I, I I I III
--
.
o
.-
I
,.
o ~rmrmml1TlllTlTfnmmITI ~r Il~-i
;0f1)(J).....0...CQt-'
(J)
Q:J0...
Q
<f1).,Q
tof1).,f1)(j)t-J.
0..CQ........
.....(j)(j)Cf1)•
,...... °i
oOilOil• I
I Ii •
-i tr
,...... -i IgjJ..1
t-'", .......1'\.)0rno'0co.......
Average. residual yieldI I I I I
0 0 0 0 0 0 0 0 0 0 0• • • • • • • • • • •0 0 0 0 0 0 0 0 0 0 00 0 0 0 0 0 0 0 0 0 0UI ~ W I'\.) t-'" 0 t-'" I'\.) W ~ Ul
0
oo
'<(J)
(J)......:Joto
4010 20 30DATA SET
70 f- I I 1 ·1 I ~ I· Io
80
Predicted, actual ask prices for bp2.str.
......Ulo
...... ...... ......I'\.) W~000
......
......o
......(() 0o 0
CDo
'..Jo0+----1---I-"""-----t---I---I----1---I-----t
......o
wo
·BOND PRICE
o>--I>N
o(f)rn--I
-0.,ro.0..t-'.
o(Tm0..•
Qo(TCQt--'
Q(j)A
-0.,1-'.
·0ro(j)
-t) .
o.,tT
-0I'\.)•3(D
0..•
......f\)•'-JU1
00C
-00:J•f\)00U1I...... -0
0 ;0• [11 ......
3DO....... 0
0 n.0.. -fID [11....... Df\) -0• ;03 ~
ID n0.. fT1•
ACTUAL ASK PRICE......oo
xx x
•• ••
- -* *x--at
•at at
•
•at
•II
......U1o
•